Checking the work

Verification habits that let you move fast without publishing fiction.

Models produce confident, fluent, occasionally false output. The professional response is neither trust nor abandonment — it is a verification habit sized to the stakes. What would be embarrassing or costly if wrong gets checked; what would be trivially wrong gets skimmed.

This is the competency that most separates people who merely use AI from people whose AI-assisted work can be relied on.

1. Stage 1 · E1 FOUNDATIONAL

   Treat output as a draft, not an answer

   Adopt one rule: nothing ships under your name unread. Statistics, quotes, names, dates, and citations are the classic failure points — models generate plausible-looking specifics precisely where evidence ran out. A precise-sounding number with a tidy attribution is a flag, not a comfort.

   For any cited claim, open the source. Confirm the figure and its context before using it. If you are on deadline and cannot check it, cut it or soften it to what you can stand behind — never publish and hope.

   Try it this week

   Take an AI draft containing at least three factual claims. Verify each against a primary source. Track your hit rate for a week — knowing your model's real error rate in your domain changes how you read it.

2. Stage 2 · E2 PROFICIENT

   Spot-check what would change the decision

   You cannot verify everything, and you do not need to. Triage: identify the claims your decision actually rests on, and check those against the original material. A 30-page summary needs its load-bearing claims verified, not its adjectives.

   Know the weak verifiers. Asking the model whether it is confident is worthless — it will say yes. Asking it to check its own work in the same thread mostly re-runs the same blind spots. Real verification is independent: the source document, fresh data, a second tool, your own recomputation.

   Try it this week

   Next summary you receive, list the three claims that would change what you do if false. Verify exactly those against the source. Note whether any failed — and what that says about the rest.

3. Stage 3 · E3 DISTINGUISHED

   Verify like an examiner

   Develop a nose for fabrication patterns: suspiciously round percentages, "widely cited" studies with no name, consensus claims ("most economists agree") that no one could have measured. When one citation out of five turns out not to exist, treat it as a signal — re-verify everything that leaned on it, because the model was inventing where evidence ran out.

   For quantitative work headed to a client, the professional standard is recompute or trace: every computed figure either re-derived from source data or traced to a source you opened. "The AI calculated it" is not a provenance.

   Try it this week

   Build a personal verification checklist for your most common deliverable: which claim types it contains, how each gets checked, and what happens to unverifiable ones (cut, soften, or attribute honestly).

4. Stage 4 · E4 EXCEPTIONAL

   Design verification into the workflow

   The exceptional operator does not check at the end — they design workflows where checking happens where errors would compound. Intermediate outputs get verified before feeding the next step; high-volume automated steps get sampled on a schedule; anything customer-visible gets a named human owner.

   They also calibrate publicly: they can tell a colleague "this model is wrong about one claim in ten in our domain, here is what that means for how we use it" — turning private instinct into a team standard.

   Try it this week

   Map your main AI workflow and mark every point where an unverified error would flow downstream. Add one checkpoint where it hurts most, and write down what "checked" concretely means there.

Ready to see where you stand? The free check scores all six competencies in about fifteen minutes.